# Repeated balanced unique-target cross-fitting sensitivity

Status: VERIFIED — independent verification 120/120 PASS

The frozen 20260717 realization remains the primary analysis. Nineteen prespecified consecutive seeds were added solely to quantify balanced-fold and fold-local training variability. Every realization contains five non-overlapping 432-target folds and one OOF score per target and model.

## Stability summary

| Quantity | Result |
|---|---:|
| Realizations | 20 |
| RMSE vs zero favorable | 20/20 |
| RMSE vs train mean favorable | 20/20 |
| Raw-cosine difference favorable | 20/20 |
| Point strict composite favorable | 20/20 |
| No negative-control point RMSE support | 20/20 |
| Pairwise fold ARI, median (range) | 0.0001 (-0.0014 to 0.0030) |

Under the prespecified interpretation rules, both RMSE contrasts and the
perturbation-specific residual cosine are `STABLE`. The raw-cosine difference is
`MIXED`: all 20 realization means are positive, but the hierarchical sensitivity
interval crosses zero. The frozen seed-20260717 conditional P value therefore
must not be presented as fold-robust directional evidence.

## Joint target + fold/training sensitivity intervals

| Metric | Mean | 95% hierarchical sensitivity interval | Between-realization SD |
|---|---:|---:|---:|
| rmse_difference_vs_zero | -0.00629188 | -0.00735218 to -0.00524033 | 0.00003336 |
| rmse_difference_vs_train_mean | -0.00046627 | -0.00070221 to -0.00022488 | 0.00002894 |
| raw_cosine_difference_vs_train_mean | 0.00122360 | -0.00021601 to 0.00249840 | 0.00027511 |
| perturbation_specific_residual_cosine | 0.04528089 | 0.02964997 to 0.06117822 | 0.00144454 |

The hierarchical interval samples a prespecified realization and then resamples targets within that realization. It is a sensitivity interval that reflects both target sampling and observed fold/training variability; it is not an independent biological-replication interval.

No cross-dataset, muscle/DMD, mechanistic, therapeutic, safety or clinical claim follows from this analysis.